Multi-prior-guided deep learning framework for hepatic vessel segmentation

Qi Zhang

doi:10.1007/s00371-026-04435-9

Multi-prior-guided deep learning framework for hepatic vessel segmentation

2026-03-22

0

PRE

AI

DOI：10.1007/s00371-026-04435-9

Original

Original request for help

Save

Abstract

En 中文

3D visualization of hepatic vessels plays a crucial role in the diagnosis and surgical planning of liver diseases, which requires the accurate segmentation of vessel structures. However, accurate segmentation of hepatic vessels remains challenging. One major challenge is inconsistent vessel enhancement, caused by variations in contrast agent administration. Another is the complex topology and abundance of fine branches in the hepatic vasculature. Numerous studies have been dedicated to this task. Yet, they often fail to capture intricate features, as they underutilize prior knowledge needed for high-fidelity visualization. In this study, we introduce the Vessel-Aware Module (VAM), comprising a Tubular Prior Fusion Block, a Neighboring Voxel Perception Block and a Projection Attention Block, to encode vessel topology, local context and spatial priors. This enhances the detection of fine vessels and generates more complete and visually coherent vascular maps. Integrating VAM into U-Mamba forms the VA-UMamba framework. Coupled with a novel Dilated Dice Loss, VA-UMamba achieves competitive performance on our high-quality dataset and two public datasets. In terms of Tree Length Detected (TD), VA-UMamba reached 74.70%, 78.72% and 78.64% on three datasets, which are 6.25%, 1.49% and 1.91% higher than the existing SOTA methods. It demonstrates the superior capability in recovering fine branches for accurate vascular visualization. The results also show significant improvements in other metrics, including Dice, Sensitivity and Hausdorff distance. Our method provides enhanced visual representations of hepatic vessels, supporting clinicians in preoperative planning through clearer anatomical insight. The code is available at: https://github.com/chen-x-h/VA-UMamba .

Keywords:

Hepatic vessel segmentation

Vascular structure analysis

Prior knowledge fusion

Multi-prior-guided deep learning framework for hepatic vessel segmentation

Abstract

Journal

Organization